Working Notes on Agent Systems/Brad Zhang

@teach_fireworks / X longform

The new idea of ​​​​OpenAI training model may be a new direction! 1. Previous dilemma:...

The new idea of ​​​​OpenAI training model may be a new direction! 1. Previous dilemma: a mess In the past, when we tried to understand the AI ​​brain, we discovered a big proble...

November 14, 2025 · 3 min read

The new idea of ​​​​OpenAI training model may be a new direction!

  • Previous dilemma: a mess In the past, when we tried to understand the AI ​​brain, we discovered a big problem: a "neuron" (thinking unit) in its "brain" would be responsible for thinking about many completely unrelated things at the same time. Metaphor: Imagine trying to listen to the "voice" of an AI neuron. What you hear instead: "Golden Retriever... ...brackets in code... ...French Revolution... ...sad emotions..." ...The wheel of a bicycle..." Result: This is a complete mess! This neuron is thinking about everything (this is technically called "superposition" or "ambiguity"), and you have no idea what it is thinking at a certain moment. It's like there are 100 people talking in a room at the same time, and you can't hear a word. Because we can't understand, AI will always be a "black box." We have no way of knowing whether it is "lying" or whether it has any dangerous ideas. This research by OpenAI found a way to "translate" this mess. They invented a tool (technically called a "sparse autoencoder") that acts like a "super translator" or "sound separator". Metaphor: You feed that "mess" of signals ("golden retriever, brackets, revolution, sadness...") to this "translator". What the translator does: It can automatically dismantle this noise and separate it into millions of very clean and single "little thoughts". Translation result: It will tell you: "Oh, in the noise just now, 5% were thinking about 'golden retrievers', 2% were thinking about 'code brackets', and 1% were thinking about 'sadness'..." These isolated, single "small thoughts" are the "Sparse Circuits" in the title of the article. This breakthrough is extremely significant because it marks the first time we have the ability to truly "open" the black box of AI and "read" its mind. Significance 1: AI security (the most important) This is a huge milestone in the field of AI security (Alignment). Previously: We have no way of knowing whether AI is “really” trying to help us or if it’s “pretending.” Now: If an AI generates a “bad idea” (such as “deceive the user,” “generate bias,” or “destroy the world”), for the first time ever, it’s possible to locate that specific “bad idea” circuit through this “translator.” It's like having brain surgery. We finally have a "brain map" that can accurately find the "brain cell" corresponding to that "bad thought", and then it is possible to "remove" it or "correct" it. Significance 2: AI Interpretability We are no longer “refining elixirs”. We’re starting to really understand why AI makes a certain decision. For example, when the AI ​​writes a poem, we can see what "little thoughts" it activates about "romance," "rhythm," or "Shakespearean style." Meaning 3: AI controllability Since we can understand it, we can control it. In the future, we may be able to "manually" turn off certain "ideas" of AI (such as bias), or "manually" activate certain "ideas" (such as "stronger creativity").

Visual summary

Article argument map

Generated from the post's content graph

FORMATTOPICCAPABILITYMARKETcoverscoverscoverssignalsFORMATarticle featureTOPICtechnical distributionTOPICretrievalTOPICtoolingMARKETopen-source builders
Mermaid outline
flowchart LR
  format-article["article feature"]
  topic-technical-distribution["technical distribution"]
  topic-retrieval["retrieval"]
  topic-tooling["tooling"]
  market-open-source-builders["open-source builders"]
  format-article -->|covers| topic-technical-distribution
  format-article -->|covers| topic-retrieval
  format-article -->|covers| topic-tooling
  format-article -->|signals| market-open-source-builders

Visual structure

Essay structure map

Built from summary and key paragraph positions

The new idea of ​​​​OpenAI training model may be a new direction! 1. Previous dilemma...THESISThe new idea of​​​​OpenAI trainingmodel may be a newdirection! 1. PreviousSIGNALThe new idea of​​​​OpenAI trainingmodel may be a newdirection!OPERATOR1. Previous dilemma: amess In the past, whenwe tried to understandthe AI ​​brain, weIMPLICATION1. Previous dilemma: amess In the past, whenwe tried to understandthe AI ​​brain, we
Mermaid outline
flowchart LR
  thesis["The new idea of ​​​​OpenAI training model may be a new direction! 1. Previous dilemma: a mess In the past,..."]
  signal["The new idea of ​​​​OpenAI training model may be a new direction!"]
  operator["1. Previous dilemma: a mess In the past, when we tried to understand the AI ​​brain, we discovered a big pr..."]
  implication["1. Previous dilemma: a mess In the past, when we tried to understand the AI ​​brain, we discovered a big pr..."]
  thesis -->|frames| signal
  signal -->|develops| operator
  operator -->|lands in| implication

Source: View the original post